I don't see an 'open()' function to override in the RichSinkFunction or the SinkFunction... so where is this open() function supposed to be?
Date: Thu, 12 Mar 2015 01:17:34 +0100 Subject: Re: Socket output stream From: [email protected] To: [email protected] Hi Emmanuel, the open() method should the right place for setting up the socket connection. It is called on the worker node before the first input arrives. Best, Fabian 2015-03-12 1:05 GMT+01:00 Emmanuel <[email protected]>: Hi Marton, Thanks for the info. I've been trying to implement a socket sink but running into 'Not Serializable' kind of issues.I was seeing in the Spark docs that this is typically an issue, where the socket should be created on the worker node, as it can't be serialized to be moved from the supervisor.http://spark.apache.org/docs/1.1.0/streaming-programming-guide.html#design-patterns-for-using-foreachrdd So, not sure how this would be implemented in Flink...My attempt (maybe very naive) looked like this: public static final class SocketSink extends RichSinkFunction<String> { private PrintWriter out; public SocketSink(String host, Integer port) throws IOException { Socket clientSocket = new Socket(host,port); out = new PrintWriter(clientSocket.getOutputStream(), true); } @Override public void invoke(String s) { out.println(s); } } maybe i should just move to Kafka directly... ;/Thanks for helpEmmanuel From: [email protected] Date: Wed, 11 Mar 2015 16:37:41 +0100 Subject: Fwd: Flink questions To: [email protected] CC: [email protected]; [email protected]; [email protected] Dear Emmanuel, I'm Marton, one of the Flink Streaming developers - Robert forwarded your issue to me. Thanks for trying out our project. 1) Debugging: TaskManager logs are currently not forwarded to the UI, but you can find them on the taskmanager machines in the log folder of your Flink distribution. We have this issue on our agenda in the very near future - they need to be accessible from the UI. 2) Output to socket: Currently we do not have a preimplemented sink for sockets (although we offer a socket source and sinks writing to Apache Kafka, Flume and RabbitMQ). You can easily implement a socket sink by extending the abstract RichSinkFunction class though. [1] For using that you can simply say dataStream.addSink(MySinkFunction()) - in that you can bring up a socket or any other service. You would create a socket in the open function and then in the invoke method you would write every value out to it. I do agree that this is a nice tool to have so I have opened a JIRA ticket for it. [2] 3) Internal data format: Robert was kind enough to offer a more detailed answer on this issue. In general streaming sinks support any file output that is supported by batch Flink including Avro. You can use this functionality by dataStream.addSink(new FileSinkFunction<>(OutputFormat)). [1] http://ci.apache.org/projects/flink/flink-docs-master/streaming_guide.html#connecting-to-the-outside-world [2] https://issues.apache.org/jira/browse/FLINK-1688 Best, MartonFrom: Emmanuel <[email protected]> Date: 11. März 2015 14:59:31 MEZ To: Robert Metzger <[email protected]>, Henry Saputra <[email protected]> Subject: Flink questions Hello, Thanks again for the help yesterday: the simple things go a long way to get me moving... I have more questions i hope I can get your opinion and input about: Debugging:What's the preferred or recommended way to proceed? I have been using some System.out.println() statements in my simple test code, and the results are confusing:First, in the UI, the logs are for the jobmanager.out, but there is never anything there; wherever i see output in a log it's on the taskmanager.out fileAlso, even more confusing is the fact that often times I just get no log at all... the UI says the topology is running, but nothing get printed out...Is there a process you'd recommend to follow to debug properly with logs? Output to socketIdeally I'd like to print out to a socket/stream and read from another machine so as not to choke the node with disk I/Os when testing performances. Not sure how to do that. Internal Data formatFinally, a practical question about data format: we ingest JSON, which is not convenient, and uses a lot of space. Internally Java/Scala prefers Tuples, and we were thinking of using ProtoBuffs. There is also Avro that could do this as I understand it... What would be the recommended way to format data internally? Thanks for your input. CheersEmmanuel
